At the Advancing AI 2024 conference in San Francisco, AMD launched its Instinct MI325X AI accelerator, directly challenging Nvidia's Blackwell chips. The new platform boasts a 1.8 times greater memory capacity, 1.3 times higher memory bandwidth, and 1.3 times the computational power compared to Nvidia's H200 HGX platform. Production is set to begin in late 2024 with shipments to customers starting in early 2025.
AMD also revealed its latest AI chip roadmap, with the CDNA 4-based MI350 series set to debut next year and the MI400 series featuring even more advanced architecture. CEO Lisa Su highlighted the expected growth of the data center, AI, and accelerator markets to $500 billion by 2028, emphasizing the role of generative AI.
MI325X employs AMD's CDNA 3 GPU architecture with 256GB of next-gen HBM3E high-bandwidth memory and 1,530 billion transistors, offering 6TB/s of memory bandwidth. Compared to Nvidia's H200, MI325X delivers 40% higher inference performance when running Meta's Llama 3.1 model.
AMD is intensifying its AI product advancements, focusing on annual iterations to effectively compete with Nvidia. MI325X succeeds the MI300X series launched last year. The company continues to improve its ROCm software, aiming to ease the migration of AI models to its chips, enhancing both inference and training by over two times with version 6.2.
Analysts note that AMD's main hurdle is gaining enterprise market share, with Nvidia's CUDA platform posing a significant barrier in the AI development space. However, AMD's new GPUs like the MI350 show improved efficiency and performance, contributing positively to its competitive positioning.
Despite challenges, AMD remains a viable alternative in the market. CEO Lisa Su has expressed confidence that the AI chip sector can support multiple successful companies without AMD needing to surpass Nvidia to achieve success.
